Give AI a useful
connection to your tools.
An MCP server exposes a defined set of tools, resources, and prompts to compatible AI applications. The protocol connects the pieces; your application still decides the permissions.
Field guide / Three useful primitives
A common language for useful capabilities.
MCP gives applications a shared way to discover and use server capabilities. Explore the different jobs of resources, tools, and prompts.
Illustrative capability examples. These are conceptual descriptions, not complete protocol messages.
THE DECISION IN VIEW
Content that supplies context.
A resource exposes content such as a document or record. The client can read that content and the host can make it available in the user’s workflow.
- Start with
- A resource identifier
- Make possible
- Content and its metadata
Example: read the current project brief from a defined resource.
01 / The short answer
An interface between an AI application and a real system.
MCP stands for Model Context Protocol. A server makes capabilities available through a shared interface so compatible applications can discover and use them. It can sit in front of a CRM, database, filesystem, or other service. It is not the model, and it does not replace the underlying business system.
Source notes
02 / Follow a request
Host → client → server → your system.
- 01AI hostThe application the person uses.
- 02MCP clientMaintains the server connection.
- 03MCP serverExposes defined capabilities.
- 04Business systemReturns data or performs an action.
Local servers can communicate through standard input and output. Remote servers can use Streamable HTTP. The right transport depends on where the application runs and who needs access.
01
Tools
Defined operations an application can invoke, such as finding a record or creating a task.
02
Resources
Context the application can read, such as documentation or a database schema.
03
Prompts
Reusable interaction templates with a defined purpose.
Source notes
From reading to doing
Follow a tool request across the boundary.
Inspect the context available to an assistant, see the work it prepares, and choose whether it may proceed. This local example illustrates application review around a connected tool; it does not send a real MCP request.
Try the workflow
03 / Place it in the stack
MCP, APIs, and tool calling solve different parts.
| Part | What it contributes | What still needs engineering |
|---|---|---|
| Business API | Operations and data in the source system | Credentials, permissions, errors, and limits |
| Tool calling | A model can request a structured action | Execution and validation of that request |
| MCP | A shared way to expose capabilities | A well-designed, authorized tool surface |
| User interface | A person can ask, inspect, and approve | Clear states, previews, and useful feedback |
04 / Make the boundary concrete
Expose the smallest useful action.
A tool called “manage everything” is difficult to understand or control. A narrower tool can have a clear purpose, required fields, and an inspectable result. Here is an illustrative tool boundary for a CRM task—not a protocol payload.
Example tool contract
- Action
prepare_follow_up_task- Required context
- Contact ID · owner · due date
- Review surface
- A preview of the task and supporting context
- Result
- A task reference or a clear, recoverable error
Follow Up Ace applies this idea to a shared assistant connected to CRM workflows. Contact priorities, reviewed outreach drafts, notes, and tasks stay close to the system the team uses.
Explore the Follow Up Ace case study →05 / Before implementation
Design the tool boundary as carefully as the screen.
List the useful outcomes
Begin with the actions a user actually needs, rather than every available API endpoint.
Separate reading from changing
Make permissions and review behavior clear for each operation.
Return evidence
Give the application identifiers, source context, and useful error information.
Test the awkward paths
Exercise missing fields, expired credentials, rate limits, duplicate requests, and partial failures.
Building the boundary correctly — permissions, review, evidence, and the awkward paths — is exactly what an MCP server development engagement covers.
Explore our MCP server development work →Keep exploring
The questions behind the question.
What is an MCP server?
A program that exposes tools, resources, and prompts to compatible AI applications through the Model Context Protocol.
Is an MCP server the same as an API?
No. It can wrap an existing API and present selected capabilities through MCP. The underlying system still performs the work.
Does an MCP server have to run in the cloud?
No. Servers can run locally or remotely, depending on the transport and deployment.
Does MCP automatically make an integration safe?
No. Authorization, validation, review, and error handling are still design and implementation responsibilities.
What is MCP server development?
Defining useful capabilities, connecting them to a real system, and implementing the permissions, validation, and recovery behavior needed for reliable use.
Why would I need an MCP server?
When you want an AI agent to take real action in your systems—updating a CRM record, querying a database, sending an email—rather than just generating text about them, instead of writing custom integration code for every model you connect.
Does ChatGPT use MCP?
Yes. OpenAI adopted MCP across its products, including the ChatGPT desktop app, in March 2025—about four months after Anthropic introduced the protocol—making it a cross-vendor standard rather than a Claude-only feature.
Is MCP the same as function calling?
No, but they work together. Function calling is the model's own ability to look at a list of tools and emit a structured request to call one; MCP is the standardized protocol that gets that tool list, and the call, between the model and whatever system performs the action.
Who created MCP and who maintains it now?
Anthropic created and open-sourced the Model Context Protocol in November 2024. In December 2025, Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation, moving it to vendor-neutral governance so its roadmap doesn't depend on any single AI company.